Your First Hour

In the paid kit, this comes right after a five-minute guided interview where Claude writes your memory file for you. Here, do a two-minute stand-in version first, then run the three proofs below on your own work.

Remember this about my business for this conversation: who I serve, how I sound
(plain, no hype, whatever fits you), my top prices or rates, and one rule I never
break (for example: never send anything without me reviewing it first).

[FILL IN YOUR ANSWERS]

Proof 1: it remembers (30 seconds)

Paste:

Introduce my business back to me the way you understood it: who we serve, how we
sound, our prices, and the one rule I gave you. Only say what I actually told you.

What you should see: your own words, recited back accurately, no re-explaining. In the full kit this holds automatically, every session, from a memory file, not a copy-paste block.

Proof 2: a real job, done your way (about 15 minutes)

Find one real email you need to answer (or one you answered recently). Paste:

Draft a reply to the email below in the voice and rules I gave you above. Verify
any number or commitment against what I told you, and flag anything you can't
verify instead of guessing. Show me the draft; don't send anything.

[PASTE THE EMAIL HERE]

What you should see: a draft that sounds like you, respects your rules (draft only, you send), and marks anything it couldn't verify instead of inventing it. In the full kit this is a saved playbook you call by name, not retyped every time.

If it misses: correct it in plain English ("shorter, never open with an apology"). When it comes back right, you've handed over your first repeat job.

Proof 3: watch a gate catch something (about 10 minutes)

Take a paragraph you already wrote and shipped: an email, a page, a post. Paste:

Run the two-gate audit below on the text under it.

GATE A - FACTS: list every number, date, name, and claim in the text. Mark each one
VERIFIED (say against what) or UNVERIFIED.

GATE B - CRAFT: check it against the voice and rules I gave you above. Flag hype,
buried points, or anything you couldn't ship as-is.

Give me the audit table, then your verdict: would this have shipped, or been held?

[PASTE THE TEXT HERE]

What you should see: an audit table of every claim, anything unsourced flagged as UNVERIFIED, and a plain verdict. Most people get at least one flag on work they already sent. That flag is the product: from now on it fires before you ship, not after.

The two gates you just used are reproduced in full below, exactly as they ship in the kit as guardrails/dual-gate-audit-loop.md. In the full kit every playbook and worker routes through this file automatically; here you ran it by hand once, on one paragraph.

The Dual-Gate Audit Loop (the core of the system)

This is what separates a governed AI team from a confident-sounding intern. Every output passes two independent gates before it reaches you. They judge different things and never cover for each other.

The loop

draft -> run Gate A and Gate B -> fix what each flags -> re-check -> ship

Repeat the deep pass only if a gate failed or a fix introduced a new problem. Cap it at 2 deep passes so it doesn't spin.

Conflict rule

When a verified fact disagrees with memory, a prior draft, or what "sounds right," the verified fact wins. "I'm not sure" is a flag to check, never a pass.

The receipt

Every audited output ends with one line so you can trust it at a glance:

Why two gates

A beautiful email with a wrong price fails Gate A. A correct price written in hype-y robot-speak fails Gate B. You need both, every time. Most AI output skips both: that's why it can't be trusted unattended.

After the hour

Pick the one job from Proof 2 you repeat most and run it three times this week until it comes back right every time. Then, and only then, add the next. One proven workflow at a time is how this scales.